Bayesian neural network-based satellite fog detection

Fog, an essential component of the hydrological cycle, is frequently experienced in North India during winter. The reduced visibility due to fog causes many accidents and delays in trains and flights, leading to loss of health and economy. Hence real-time detection and forecast of fog are crucial for mitigating these losses. The study proposes an algorithm to detect fog using satellite observations. The algorithm consists of Bayesian Neural Networks containing weights as probability distributions, unlike ordinary neural networks that treat weights as deterministic parameters. This algorithm provides prediction uncertainty. Both epistemic (data-dependent) and aleatoric (model-dependent) uncertainties are modelled. The final output is the percentage chances of fog which can be suitably thresholded into fog/no-fog. In this study, in situ airport weather records (METAR) are used as reference observations, whereas satellite observations are obtained from the 6 bands of the INSAT-3D geostationary satellite (with a spatial resolution of 4 km). Sub-hourly data of wintertime observations from 2016 to 2020 for seven cities spread across North India are used to train and validate the proposed methodology. The model performs better than the INSAT-3D fog product developed by ISRO. The critical success index of INSAT-3D fog product and the proposed method are 0.17 and 0.44, respectively, whereas Cohen’s Kappa values are 0.22 and 0.50, respectively. The uncertainty analysis shows that aleatoric uncertainty is generally higher than epistemic uncertainty. Moreover, for observations having higher aleatoric uncertainty, the epistemic uncertainty is also high, showing a positive correlation. The real-time predictions are disseminated on the website (www.fog.iitk.ac.in) for the public and scientists. This work is a part of the Fog Prediction using Data Science project.